Ultimately, of course, the Cloud is only as good as the bits it is constructed from, a situation which creates something of a Curate’s Egg for users. The conceptual and operational liberty that the Cloud brings to the development of new business models and systems, particularly through the decoupling possible between the strictures of specific technology stacks and the operational processes that run on them, has to be balanced against what the major technology developers feel able or willing to offer us.
So last week’s Intel Developer Forum was always going to be important in setting the scene for future Cloud developments – or at least the developments Intel would like us to adopt and run with.
Cloud developments did, of course, only occupy a corner of what this mainly semiconductor technology event covered. But the trends the company identifies moving forward are important, if only for reinforcing the breadth of the cloudscape those responsible for implementing Cloud services have to consider. This the company segmented into three broad categories: security, power, and the client-side.
Though not identified by the company directly, much of its view of security issues are being driven by the Security-By-Policy model. Here, tightly defined and administered security policies are seen as a far better approach than inserting piecemeal security tools into a Cloud infrastructure.
Underpinning this approach will be comprehensive monitoring tools, and analysis of the resulting security and operational monitoring data is seen by Intel as being huge. This will need not only real-time monitoring of what is happening but real-time monitoring of the monitoring and sensing systems themselves for deviations that might indicate the sensors are under attack or have been compromised.
This maps onto the security-by-policy which, given the growing number of possible points of weakness is fast becoming the best way to manage the security requirements of Cloud-based services. As an example of this problem Intel itself points out that each new sensor is a potential security vulnerability, including GPS, accelerometers, gyroscopes and compasses to name but a few new options.
Such inbound data will need to be handled in real time, being run through algorithms that can handle unstructured data with no schema or consistent syntax and semantics. This obviously poses some significant problems in terms of both management and use of compute resources.
It would at this point be normal to assume that this would mean the use of large numbers of high-powered servers running the biggest Xeon processors. But in practice that may not be the case.
I have written before about Intel talking up what it calls heterogeneous processing – the combination of different processor cores on a single chip. This security-by-policy model could provide an ideal reason to start building them. The common mix identified by Intel in the past would include one or possibly two Xeon devices coupled on a single chip with a large number of low-powered, averagely performing Atom processors or GPU (Graphics Processor Unit) devices. Such an architecture could be used to create a grid-based processing environment where the Xeon(s) managed the distribution of real-time analytical tasks between the Atom and GPU processors. GPUs in particular are seen as well placed to work as analytical processors in such a role. The Xeons could then also manage the collation of the results.
This would, in effect, be the basis of a parallel compute, security policy server where the performance and capability was determined by the number of heterogeneous cores in the server design.
When it comes to power issues Intel is aiming to optimise offerings at every level in the datacentre. This does, however, beg some questions.
For example, in an increasingly commoditized Cloud environment the issues are less about the technology per se and rather about the overall contribution to the desired end-result. Intel champions cost and efficiency, and they are important criteria, but these days efficiency can often mean `good enough’ on an all-round basis where technology plays only a minor part.
Low cost, reasonably performing, Linux running systems running manycore ARM processors, for example, offer an arguably more efficient platform overall. – and it is the direction some of the biggest service providers are looking. The key here is that decoupling of the technology stack and the operational processes being run.
And as those processes get richer and more self-contained, efficiency is likely to come down to a combination of power consumption, bandwidth and manageability.
Power consumption points squarely at devices such as ARM and Intel’s Atom processors – with ARM seeming to lead the charge at the moment, though Atom is starting to gain a bit more traction which will only be helped by the move of Google to port the Android operating system to it. Bandwidth also points that way. GPUs are also gaining traction where a combination of low power consumption and high compute performance are important criteria.
The raw bandwidth of the network is growing, but the bottlenecks from an infrastructure point of view are found in reading in, and delivering out, the vast amounts of data these increasingly rich processes will require. Large volumes of low cost, low power servers are generally seen as the favoured approach here, so manycore ARM processors running Linux-based environments would seem to hold first place for now, though to be fair this race is only on its opening laps.
In the power arena, manageability is once again likely to prove to be one area where Intel can maintain some serious traction with the heterogeneous processor approach. A Xeon core could provide the workload management for a large number of smaller processors, dividing up tasks across them as required and then collating the results. This in turn could be used to create an environment in which tasks can be automatically and dynamically allocated and reallocated across the resources of a datacentre in accordance with policies that define operational priorities and manage micro-billing systems. They would use algorithms that balance the policies against actual and predicted workloads, resource availability and a host of other real time monitored processes.
Like all the big vendors with roots deep into the traditional business models Intel still shows signs of wanting to find some way of locking users in to its own technology, even though this runs counter to much of what the Cloud stands for, and is definitely counter to the users’ best benefits.
This is particularly noticeable at the client end of the Cloud spectrum, where the company’s tactic is the development of the `client-aware cloud’. Here, every time a client device logs on to the Cloud it will deliver information about itself to the host systems. That could range from basic information about the bandwidth it can send or receive, or the format required when applications deliver services, through to operational security information such as the specific service authentications it has available. Such information could be useful in allowing servers to rapidly configure themselves to accommodate that client’s specific requirements
It could also give an excellent opportunity for exploiting device-to-device recognition and providing consequent operational advantages. Given that Intel is offering processors with V-Pro, a system that allows remote access regardless of the state of the machine, a V-Pro to V-Pro connection between a client and a server could be something worth exploiting.
This, of course, is driven by Intel’s fears about thin clients – a market that is growing fast and has huge potential as a core component of Cloud service delivery. It is also a market where Intel does not `dominate’.
The company is still a fan of the fat client – a.k.a. the traditional PC – and there is still an important role for such machines. As I have written before, many of the `thin client’ devices seem like essential fashion accessories to be seen with, but when it comes to doing real work are often set aside in favour of a laptop PC. But that situation will change as what constitutes a `thin’ client changes to a range of `right-sized’ clients.
It will also change as the availability of remote access to the Cloud becomes both more ubiquitous and more reliable. Then, the right-sized, always-connected client will likely be the death knell of the fat client PC.
For now however, a simple rule of thumb can be used: a fat client equals any system required to be used when not connected. For now that still represents a large number of situations. Ironically, despite the fact that online availability is becoming available on trains, planes and the rest, I suspect it will be a while before this stops being premium rate delivery mechanisms where hardly any economy-class users will bother and many company CFOs will find it gives them cause to study executive expenses claims a little too closely.
Given some recent famous, and widespread, Cloud service outages anyone who wants to ensure that their input is at least stored somewhere will still want a fat client. Then what they will want is fat bandwidth for uploads and downloads when a connection is required.
But Intel is wrong to think it will continue to be the huge market it once was. As a writer my connection with the Cloud can be sporadic – in fact there are times I require it that way so I am not disturbed. I am sure there are many others with a similar requirement. But I am aware I am in a minority and, as the broad spectrum of users grows, that minority will shrink. For Intel to hang its hat too firmly on the fat client will be extremely short-sighted.
Instead it should be promoting the widest possible agnostic connectivity. It should remember that, though it currently stands at a position of great strength in the world of commodity servers for datacentres, the decoupling of operational process from technology stack means that this market can – and probably will - change just as quickly as fat to thin clients.
Today it would be unthinkable to suggest that a small, manycore processor maker like Tilera could possibly take serious marketshare from Intel in the server space. But with 100-core devices running Linux at high speed and low power – and devices that can be readily substituted for Intel chips in existing server designs - who is to say that it might be quite an acceptable notion to the large service providers this time next year.
